The present application relates to a data processing device, a data processing method, and a program, and particularly to a data processing device, a data processing method, and a program that make it possible to establish a method which facilitates the calculation and adjustment of a parameter and which eliminates a need for an advance database in a case of estimating an electric apparatus from current information obtained.
Techniques of estimating an electric apparatus connected beyond a distribution board from information on current measured by the distribution board are referred to as non-intrusive load monitoring (hereinafter referred to as NILM), and have been studied since the 1980s. NILM has great advantages of not requiring a measuring instrument for each of individual electric apparatuses (loads) and being able to grasp the states of all the electric apparatuses connected beyond one point on the basis of only a result of measurement at the one point.
As a representative technique of NILM, U.S. Pat. No. 4,858,141 (hereinafter referred to as Patent Document 1), for example, discloses a technique of identifying an electric apparatus by calculating real power and reactive power from measurements of current and voltage and clustering amounts of change in the real power and the reactive power. The amounts of change are obtained because the real power and the reactive power being measured change when the electric apparatus is turned on and off.
FIG. 1 is a diagram shown as FIG. 8 in Patent Document 1. In FIG. 1, real power and reactive power when a refrigerator and a heater are on and off are plotted on a two-dimensional plane having the real power and the reactive power as axes thereof. FIG. 1 shows that the on and off states of the electric apparatuses are plotted at positions symmetric with respect to a point.
The method of Patent Document 1 obtains difference at the on and off times, and thus uses only information on the moments of changes. In addition, a change point detector (change detector) is necessary, and when the change point detector fails (misses detecting turn-on or turn-off, or excessively makes erroneous detection of changes), an entire process in a subsequent stage fails.
That is, the method of Patent Document 1 has the following problems. First, because difference at on and off times is obtained, only information on the moments of changes is used. Second, it is difficult to adjust a threshold value for change point detection, and when a change point detector (net change detection) fails, an entire process in a subsequent stage fails. Third, while the method of Patent Document 1 was able to be applied because many household electric appliances in the 1980s were simple loads, states of many modern electric apparatuses cannot be classified into an on state and an off state alone, so that the method of Patent Document 1 does not function well.
In order to deal with recent electric apparatuses that consume power in a complex manner, there has arisen a need to perform some complex process also on the side of NILM. As attempts to meet the need, many methods using a discriminative model (discriminant model, classification) have been proposed. There are for example methods disclosed in Japanese Patent Laid-Open No. 2001-330630 and PCT Patent Publication No. WO01/077696 (hereinafter referred to as Patent Documents 2 and 3) as methods using an LMC (Large Margin Classifier) such as a support vector machine or the like for a discriminative model.
A discriminative model such as Adaboost, a support vector machine, or the like is known to exhibit very high discriminating performance when a feature quantity is selected well and there is a sufficiently large amount of sample data for learning. This method may therefore be considered to be effective in improving accuracy. On the other hand, however, in the case of a discriminative model, unlike a generative model such as an HMM or the like, it is necessary to prepare learning data in advance and complete learning, and it is further necessary to retain results of the learning as a database. That is, there is a disadvantage in that unknown electric apparatuses cannot be handled.
There are techniques for making discrimination by a simple linear model as in Japanese Patent Laid-Open No. 2008-039492 (hereinafter referred to as Patent Document 4) and Shinkichi Inagaki, Tsukasa Egami, Tatsuya Suzuki, Hisahide Nakamura, and Koichi Ito, “Non-Intrusive Type Operation State Monitoring System for Electric Apparatuses—Solution Based on Integer Programming with Attention Directed to Discrete States of Operation—,” Proceedings of 42th Workshop on Discrete Event Systems, The Society of Instrument and Control Engineers, pp. 33-38, 2007 (hereinafter referred to as Non-Patent Document 1). However, the techniques have the same problems as the techniques disclosed in Patent Documents 2 and 3 in that an advance database is necessary. Other methods using a database prepared in advance include methods proposed as Japanese Patent Laid-Open Nos. 2006-017456 and 2009-257952 (hereinafter referred to as Patent Documents 5 and 6).
After all, the above-described methods in the past involve a tradeoff between accuracy and a method not requiring advance registration. Household electric apparatuses have recently been greatly diversified, and a discriminative model that needs advance learning is considered to be effectively unfit for household use. A method not requiring advance registration is therefore more desirable.
Accordingly, attempts to use a generative model rather than a discriminative model that needs advance learning have already been made. For example, techniques in which a hidden Markov model (HMM) is applied as a generative model have been proposed (see Bons M., Deville Y., Schang D. 1994. Non-intrusive electrical load monitoring using Hidden Markov Models. Third International Energy Efficiency and DSM Conference, October 31, Vancouver, Canada., p. 7 (Non-Patent Document 2) and Hisahide Nakamura, Koichi Ito, and Tatsuya Suzuki, “Electric Apparatus Operation Condition Monitoring System Based on Hidden Markov Model,” IEEJ Transactions on Power and Energy, Vol. 126, No. 12, pp. 1223-1229, 2006 (Non-Patent Document 3), for example).
However, in a case where a simple HMM is applied as a generative model, the number of states explodes (becomes enormous) as the number of electric apparatuses is increased, and a practical system cannot be constructed. For example, supposing that each electric apparatus has two on and off states, and that the number of electric apparatuses is n, a necessary number of states is 2n. Further, the size of state transition probability is the square of 2n (2n)2. Supposing that there are 20 electric apparatuses in total in an ordinary household (which are by no means a large number in recent years), a necessary number of states is 220=1,048,576, and the size of state transition probability is 1,099,511,627,776. This size is on the order of one terabit, and is difficult to be handled even by personal computers in recent years.
Incidentally, the method of Patent Document 1 is also based on clustering, and can be considered to be a primitive generative model, so that advance registration is not necessary. A method for solving a problem without modeling in an era before an approach based on a stochastic model became common, such as the method of Patent Document 1, is referred to as a heuristic method. A heuristic method may be useful as a first step, but has problems in that parameters such as a threshold value and the like increase rapidly as the method is extended, and in that it becomes difficult to adjust the parameters.
An automatic recognition technology using computers has recently become able to solve various difficult problems as a result of the introduction of a stochastic model. When modeling can be performed well by a stochastic model, most parameters can be obtained by maximum likelihood estimation (ML estimation, Maximum Likelihood), posterior probability maximization (MAP estimation, Maximum A Posteriori), a minimum classification error (MCE), and the like. The use of a discriminative model such as a support vector machine or the like and a generative model such as an HMM or the like corresponds to modeling by a stochastic model.